Open dshemetov opened 3 months ago
Aside: if we are super concerned with space (not clear we are at the moment, seems a "nice to have" rather than a "mandatory for adding new features"), we may want to investigate ways to use {butcher} for existing workflows.
{butcher}
Originally posted by @dajmcdon in https://github.com/cmu-delphi/epipredict/issues/293#issuecomment-1936620136
A bit strange that the lm fit object here is way larger than the training dataset. I'm guessing there's per-quantile duplication going on here.
``` r library(epipredict) #> Loading required package: epiprocess #> #> Attaching package: 'epiprocess' #> The following object is masked from 'package:stats': #> #> filter #> Loading required package: parsnip # Basic fitting example jhu <- case_death_rate_subset r <- epi_recipe(jhu) %>% step_epi_lag(death_rate, lag = c(0, 7, 14)) %>% step_epi_ahead(death_rate, ahead = 7) %>% step_epi_lag(case_rate, lag = c(0, 7, 14)) %>% step_epi_naomit() wf <- epi_workflow(r, parsnip::linear_reg()) %>% fit(jhu) latest <- jhu %>% dplyr::filter(time_value >= max(time_value) - 14) preds <- predict(wf, latest) # Recursively apply a function and flatten the result apply_nested_flatten <- function(nested_list, func, depth = 1) { lapply(nested_list, function(item) { if (is.list(item) && depth > 1) { apply_nested_flatten(item, func, depth - 1) } else { func(item) } }) %>% `names<-`(names(nested_list)) %>% purrr::list_flatten() } # Get a sense for the object sizes lobstr::obj_sizes(!!!list(jhu = jhu, r = r, wf = wf, latest = latest, preds = preds)) #> jhu : 661.51 kB #> r : 16.74 kB #> wf : 4.55 MB #> latest: 27.70 kB #> preds : 2.26 kB # Inspect the wf object, note that the fit object is the largest lobstr::obj_sizes(!!!apply_nested_flatten(wf, function(x) x, 3)) #> pre_actions_recipe : 20.75 kB #> pre_mold_predictors : 928.59 kB #> pre_mold_outcomes : 155.10 kB #> pre_mold_blueprint : 14.94 kB #> pre_mold_extras_roles : 622.83 kB #> pre_case_weights : 0 B #> fit_actions_model : 2.58 kB #> fit_fit_lvl : 0 B #> fit_fit_spec : 19.69 kB #> fit_fit_fit : 2.80 MB #> fit_fit_preproc_x_var : 0 B #> fit_fit_preproc_y_var : 0 B #> fit_fit_elapsed_elapsed: 56 B #> fit_meta_max_time_value: 112 B #> fit_meta_as_of : 0 B #> trained : 0 B # Go deeper lobstr::obj_sizes(!!!apply_nested_flatten(wf, function(x) x, 4)) #> pre_actions_recipe_recipe : 18.54 kB #> pre_actions_recipe_blueprint : 1.65 kB #> pre_mold_predictors_lag_0_death_rate : 154.61 kB #> pre_mold_predictors_lag_7_death_rate : 154.61 kB #> pre_mold_predictors_lag_14_death_rate : 154.61 kB #> pre_mold_predictors_lag_0_case_rate : 154.61 kB #> pre_mold_predictors_lag_7_case_rate : 154.61 kB #> pre_mold_predictors_lag_14_case_rate : 154.61 kB #> pre_mold_outcomes_ahead_7_death_rate : 154.61 kB #> pre_mold_blueprint_intercept : 0 B #> pre_mold_blueprint_allow_novel_levels : 0 B #> pre_mold_blueprint_composition : 0 B #> pre_mold_blueprint_ptypes_predictors : 392 B #> pre_mold_blueprint_ptypes_outcomes : 296 B #> pre_mold_blueprint_fresh : 0 B #> pre_mold_blueprint_strings_as_factors : 0 B #> pre_mold_blueprint_recipe : 12.58 kB #> pre_mold_blueprint_extra_role_ptypes_time_value: 464 B #> pre_mold_blueprint_extra_role_ptypes_geo_value : 408 B #> pre_mold_blueprint_extra_role_ptypes_raw : 472 B #> pre_mold_extras_roles_time_value : 155.08 kB #> pre_mold_extras_roles_geo_value : 157.97 kB #> pre_mold_extras_roles_raw : 309.57 kB #> pre_case_weights : 0 B #> fit_actions_model_spec : 2.08 kB #> fit_actions_model_formula : 0 B #> fit_fit_lvl : 0 B #> fit_fit_spec_args_penalty : 0 B #> fit_fit_spec_args_mixture : 0 B #> fit_fit_spec_mode : 0 B #> fit_fit_spec_user_specified_mode : 0 B #> fit_fit_spec_method_libs : 112 B #> fit_fit_spec_method_fit : 1.52 kB #> fit_fit_spec_method_pred : 17.37 kB #> fit_fit_spec_engine : 0 B #> fit_fit_spec_user_specified_engine : 0 B #> fit_fit_fit_coefficients : 344 B #> fit_fit_fit_residuals : 155.98 kB #> fit_fit_fit_effects : 309.27 kB #> fit_fit_fit_rank : 56 B #> fit_fit_fit_fitted.values : 154.66 kB #> fit_fit_fit_assign : 80 B #> fit_fit_fit_qr : 1.08 MB #> fit_fit_fit_df.residual : 56 B #> fit_fit_fit_call : 7.92 kB #> fit_fit_fit_terms : 4.55 kB #> fit_fit_fit_model : 1.08 MB #> fit_fit_preproc_x_var : 0 B #> fit_fit_preproc_y_var : 56 B #> fit_fit_elapsed_elapsed : 56 B #> fit_meta_max_time_value : 112 B #> fit_meta_as_of : 0 B #> trained : 0 B # Use butcher to reduce the memory small_lm <- butcher::butcher(wf$fit$fit$fit, verbose = TRUE) #> ✔ Memory released: 1.24 MB #> ✖ Disabled: `print()`, `summary()`, and `fitted()` butcher::weigh(wf$fit$fit$fit) %>% print(n=20) #> # A tibble: 21 × 2 #> object size #> <chr> <dbl> #> 1 terms 1.10 #> 2 call 1.09 #> 3 qr.qr 1.08 #> 4 effects 0.310 #> 5 residuals 0.156 #> 6 fitted.values 0.156 #> 7 model...y 0.155 #> 8 model.lag_0_death_rate 0.155 #> 9 model.lag_7_death_rate 0.155 #> 10 model.lag_14_death_rate 0.155 #> 11 model.lag_0_case_rate 0.155 #> 12 model.lag_7_case_rate 0.155 #> 13 model.lag_14_case_rate 0.155 #> 14 coefficients 0.000848 #> 15 qr.qraux 0.000112 #> 16 assign 0.00008 #> 17 qr.pivot 0.00008 #> 18 rank 0.000056 #> 19 qr.tol 0.000056 #> 20 qr.rank 0.000056 #> # ℹ 1 more row # Still a lot of memory used, even after butcher's cleanup butcher::weigh(small_lm) %>% print(n=20) #> # A tibble: 21 × 2 #> object size #> <chr> <dbl> #> 1 qr.qr 1.08 #> 2 effects 0.310 #> 3 residuals 0.156 #> 4 model...y 0.155 #> 5 model.lag_0_death_rate 0.155 #> 6 model.lag_7_death_rate 0.155 #> 7 model.lag_14_death_rate 0.155 #> 8 model.lag_0_case_rate 0.155 #> 9 model.lag_7_case_rate 0.155 #> 10 model.lag_14_case_rate 0.155 #> 11 terms 0.00570 #> 12 coefficients 0.000848 #> 13 qr.qraux 0.000112 #> 14 call 0.000112 #> 15 assign 0.00008 #> 16 qr.pivot 0.00008 #> 17 rank 0.000056 #> 18 qr.tol 0.000056 #> 19 qr.rank 0.000056 #> 20 df.residual 0.000056 #> # ℹ 1 more row
Created on 2024-03-28 with reprex v2.0.2
Aside: if we are super concerned with space (not clear we are at the moment, seems a "nice to have" rather than a "mandatory for adding new features"), we may want to investigate ways to use
{butcher}for existing workflows.Originally posted by @dajmcdon in https://github.com/cmu-delphi/epipredict/issues/293#issuecomment-1936620136
A bit strange that the lm fit object here is way larger than the training dataset. I'm guessing there's per-quantile duplication going on here.
Created on 2024-03-28 with reprex v2.0.2